Parallel performance is measured relative to a baseline, not by processor count alone.

If a workload takes time $T_1$ on one processing resource and $T_p$ on $p$ resources, its speedup is

$$S_p=\frac{T_1}{T_p}.$$

Its parallel efficiency is

$$E_p=\frac{S_p}{p}.$$

For example, if a computation falls from 100 s to 16 s on 8 cores,

$$S_8=6.25,\qquad E_8\approx0.78.$$

The missing efficiency can come from serial work, synchronization, communication, load imbalance, memory contention and the overhead of creating or scheduling parallel work.

Strong scaling keeps the problem size fixed while increasing resources. Eventually there is too little useful work per resource and speedup saturates.

Weak scaling increases problem size with resource count so each resource receives roughly constant useful work. It asks whether a larger machine can solve a proportionally larger problem in comparable time.

A parallel design should therefore be evaluated by measured time and scaling behavior, not by how many threads or devices it activates.